Upstage releases open-weight Solar Open2 250B
Upstage has released Solar Open2 250B on Hugging Face, an open-weight model with 250 billion total parameters and support for context lengths up to 1 million tokens. Utilizing a Hybrid-Attention Mixture-of-Experts architecture that activates 15 billion parameters per token, it is specialized for AI agent capabilities, tool calling, and document processing.
Upstage effectively tackles long-context inference constraints by pairing sparse MoE routing with hybrid linear attention.
- –Activates just 15B parameters per token, significantly reducing inference latency and hardware cost.
- –Hybrid-attention architecture enables efficient 1M-token context retention without quadratic memory bottlenecks.
- –Purpose-built training in simulated office environments gives it strong multi-lingual document and tool execution performance.
DISCOVERED
1d ago
2026-07-22
PUBLISHED
1d ago
2026-07-22
RELEVANCE
AUTHOR
_akhaliq